Two Flavors are Better than One: Socket A

First in line today is the socket A Sempron based on the K7 architecture.



Socket A Sempron 2800+


If the familiar rectangular shape of the core looks a little too familiar, that's probably for very good reason. The K7 Sempron is actually a 166MHz FSB 256KB Thoroughbred-B core.



Sempron's Secret


Upon first learning this, we were a little worried that we would see overlapping parts where new Semprons would actually be old Athlon XPs. Fortunately, this is not the case.

The slowest Athlon XP 166MHz FSB Thoroughbred-B was the 2600+ with a 12.5 multiplier (2.083GHz core clock). The highest K7 Sempron clock speed is 2GHz with a 12 multiplier.

What this does mean is that higher model numbered Semprons will be slower than "equivalent" Athlon XP processors. This does make a little bit of sense in that Sempron denotes a budget processor, which will be slower than the mainstream line. This is the same type of thing that we see when looking at a 2.8 GHz Pentium 4 versus a 2.8GHz Celeron. They have the same "speed", but different performance.

In our opinion, this is a problem with marketing by clock speed, and we are disappointed to see AMD recreating a problem with artificial model numbers. So now, we have 4 different 2800+ CPUs (AXP Barton, AXP Tbred, Sempron Tbred, A64 Newcastle) and really no reason why this situation should have to exist. Hopefully, future model numbering plans will be better able to handle situations such as this.

Index Two Flavors are Better than One: Socket 754
Comments Locked

55 Comments

View All Comments

  • chuwawa - Thursday, July 29, 2004 - link

    Great article! I'm looking forward to the overclocking one that's to follow. Things I'd like to see compared in it:

    1)AthlonXP-m at 2.4ghz
    2)Athlon64 3000+ at 2.4ghz
    3)push the Sempron 3100+ as high as it would go without getting too hot with the retail heatsink
    4)Do #3 with a good aftermarket heatsink

    Probably do 2-4 with one of the cheaper mobo's like the Chaintech one (it's a budget CPU afterall).
    Perhaps also do it with the $150 Epox board to see if there's a large difference, but if it's a large hassle, it's no big deal.

    One thing that's bothering me is this mentality that s754 is dead and that s939 has the only viable upgrade path. How are people so sure without a doubt? If you go s754 now, you'll be able to upgrade to the 64 3700+ which is a pretty decent processor.

    From what I understand, AMD is about to move into dual-core processors in about a year and won't those need a new socket to function?


  • KF - Thursday, July 29, 2004 - link

    The Sempron line turns out to be interesting to the performance-minded after all.

    Now that AMD is scaling the performance rating/model number of their value line to Intel Celeron Ds, Intel zig-zags again. Why would Intel marketing permit consumers to make an easy, obvious and sound judgement, especially considering which way it would go?

    >You don't mean to tell me you still believe AMD's line about "PR rating
    >is scaled off the old thunderbird" do ya.:P

    Since that is the only reasonable understanding, a reasonable person should. I didn't see TrogdorJW getting into this at all though.

    Actually, the reference to TBird scaling disappeared off the AMD site maybe 2 years ago. At least I haven't been able to find it for years. What they did have (and somewhat hard to find) was a paper on a performance-based rating which AMD proposed. Surely no one at AMD truly believed Intel would ever go along, did they? What they don't have in the paper is what the base processor (the one you scale everything to) would be. Here's how you do it (in the paper.) You do the group of benchmarks and then do an averaging. AMD calls their averaging unweighted, which really means the weightings are 1.(If you don't pick a 1 weighting, then people will argue over the weighting factor.) Then you scale the average to a reference processor's rating. That could give you a ratio or per cent, but instead you could take the clock speed of one reference processor, to come up with a imaginary "equivalent" clock. Of course, true speed even within the same chip line doesn't scale linearly to clock speed; it is just pretty close when the clocks are pretty close; which makes the equivalence imaginary.

    One problem is that the benchmarks that have the same name get updated by the creators at intervals. They might not have the same ingredients, or at least they have the newer version of the old programs. (QuakeIII vs QuakeII?) Should you keep the old benchmarks, which would provide a constant standard of measure, or if not, when should you phase in the new? If you change the base benchmark, then you no longer have an absolute scale. OTOH, if you don't change, you have a set of benchmarks which get further and further away from real-life performance. There is no way to settle this abstractly. As is well known, when Intel changes processor generations, programs get more optimized to the new processors on a timeline of a year or two. Do you go back and change the model numbers of the old models, now that everything on which it is based is changed? This amounts to the result that a rating (model number) which was conservative to start with gets to looking pretty dubious after 1-2 years. Right now Sempron-to-Celeron ratings look VERY conservative.The initial XP ratings were very conservative. People are satisfied when expectations are exceeded. They think that is as it should be. They get irrate if it falls short. Therefore you have to underrate to stay on people's good side.

    I can see how a pure model number would be OK in the server market, where the buyers are likely technically minded and can figure it out on their own. And I can see how Intel could get away with it even with consumer chips; a higher model number is better, and worth more, even if you can't tell how much. Vaguarities can be worked to advantage. Salesmanship can make a big deal out of tiny differences, and "pump the chump" for more dough. There would be dismay anyway as Intel reverts to higher instruction-per-cycle chips for the next generation, and maybe doing this asynchronously will cushion that. If you do it simultaneously, you arouse suspicion. Intel will probably hand out brochures with there own "Intelmark" ratings during the transition, like they did in the 486-to-Pentium transition.

    I think it is obvious that AMD picked the Tbird to scale the model number to because it was just about the same as the Pentium II/IIIs. That was OK while programs were still optimizing for that chip. As companies optimized for the various, drastcally different P4's, the connection to real-life performance of AMDs rating got pretty tenuous. They seem to have resynced this with A64 chips, but they aren't saying to what AFAIK. They are being vague about Sempron. Of course AMD marketing expects the rating to be compared to Intel chips. What else would people do? I have no argument with the marketing necessity of this for AMD. They are distinctly behind Intel in marketing savy anyway, and IAC AMD can't market the same way as the leader does. But AMD should have all the details spelled out somewhere for the technically and scientificly minded. These are the people that are (in effect) making or breaking AMDs reputation in the media. Getting on the good side of these people did wonders for the original K7. That, and a great chip too.

    -------------
    >Along with the added reliability Intel offers
    Intel doesn't offer any added realiability. About the only time you get into this is when you buy some oddball "blowout" revision on some closeout site to save $3. Sometimes I wonder about the sanity of system builders who do this and lose the bet.

    ---------------

    >the only real reason AMD got into the chipset business in the first place was because ...

    There was another, rather more difficult to resolve, reason. To incease market penetration, Socket 7 chipsets were never restricted by Intel (there were a skillion makers), while Intel reversed course and denied licenses to the PII bus relentlessly, with the idea of heading off possible AMD PII clones (when AMD bought NexGen.) Intel had terminated (in court) any further cross-licensing to AMD, with the intent of shutting down clones from AMD. (Years afterward, they formed a peace treaty out of necessity.) VIA only lefthandly, and against Intel's determination, got a license by buying Cyrix from National Semiconductor. NS was unloading Cyrix for practially nothing after canceling whatever future plans it had, having effectively destroyed Cyrix in the interim. And Cyrix only got a license from Intel to settle a patent suit Intel was losing. Intel fought VIA in court but did not prevail. Intel never intended to let anyone get a license until AMD was kaput. Unless AMD was willing to develop a chipset for the K7, and give the rights away, there would be no K7. Rather than concede, AMDs solution was to license the EV bus from DEC (which was beginning to slide), but use connectors similar or identical to whatever Intel was using (to keep the product cheap.) Even with that, VIA directed its efforts to PII chipsets, and was very late for K7, until they realized what a smash hit the K7 was. None of AMDs "partners" ever sacificed or risked anything on AMDs behalf. They just cleaned up on eaay pickin's. The only reason Nvidia produced chipsets for AMD is that Intel to this day won't let them license for Intel's buses, other than within the MS XBox, which is what the Nvidia chipset was actually developed for. What would you rather do: Make a chipset for 20% of the market or 80%. It (apparently) is AMD's chipset design that makes the adaptation easy.


  • thebluesgnr - Thursday, July 29, 2004 - link

    AMD created a chipset for Athlon 64 (AMD8151) but they licensed it to ALI, which released it as ALI M1687.
  • Zebo - Thursday, July 29, 2004 - link

    Trogder,

    There is no confusion in my mind.

    Value
    --------
    The sempron line, the 3100+, is meant to say = 3.1Ghz CeleronD's performance.

    Premium
    ---------
    The A64 line, the 2800+, is meant to say = 2.8Ghz P4C/E performance.

    You don't mean to tell me you still believe AMD's line about "PR rating is scaled off the old thunderbird" do ya.:P
  • TrogdorJW - Wednesday, July 28, 2004 - link

    Derek, things I would like to see in the OC update? Ah, LOTS of stuff.

    First, we really need a higher-end Athlon XP represented in the charts. Given the costs of the mobile parts and the number of enthusiasts that overclock them, at the very least a 2.2/3200+ would be a nice addition, and a 2.4 GHz Barton core would be good as well.

    Going along with that, some OC values from the Celeron D would be good to include in the graphs (not the 775-LGA versions, though - the boards are far too expensive to be used in a "budget" system).

    How about also throwing in the the other speed Spempron and Celeron D chips?

    Another addition that I would like would be to actually list the clock speeds of all CPUs in parantheses (since the model numbers basically just serve to obfuscate that knowledge these days). While we should all know (hopefully) that the Sempron 3100+ is a 1.8 GHz part and the A64 2800+ is 1.8 GHz part, the more CPUs that are listed in a benchmark, the harder it's going to be to remember what's what.

    Incidentally, your article was spot on in saying that the new processor ratings are garbage. Now that Intel has sort of followed AMD into the model number arena, I would like to see AMD drop the performance ratings (since they don't mean squat) and shift to model numbers like they have on the Opteron and FX parts. The overlapping Athlon FX, Athlon 64, and now Sempron names is going to really cause confusion among the less-informed public (which is maybe the whole point?)

    Last request for the OC article: how about some OC attempts with the retail 3100+ heatsink as well as with an aftermarket heatsink? Since it's still a budget processor, a lot of people might want to just buy the retail version and overclock that a little bit instead of buying an extra $50 HSF.
  • TrogdorJW - Wednesday, July 28, 2004 - link

    #39 - That would be the AMD750/751 chipset (at least for the earlier Athlon CPUs). When AMD first released the Athlon, and later with the switch to socket A, they created a decent chipset so that they wouldn't be dependent on the likes of Via, SiS, and ALi to provide quality chipsets. Hell, the biggest problems with the good old Super 7 CPUs from AMD (K6-2 and K6-III) was the POS motherboards and chipsets!

    Luckily, the entry of Nvidia into the chipset market has given Via some incentive to not hold back (or delay) AMD chipsets, and now AMD has more or less exited the chipset business as their partners are doing a good job. I should note that the only real reason AMD got into the chipset business in the first place was because they had such a small market share that no one was concentrating on producing quality chipsets for them. It was always Intel systems first, AMD a distant second. The quality of the Athlon architecture helped to reverse that trend, and now there is a good market for AMD system chipsets.
  • draco8099 - Wednesday, July 28, 2004 - link

    #36

    First the memory controller in the Athlon64 family is on the fricking CPU, they still need other chips on the mother board to run things.

    Yes, AMD does make a chipset for the Opteron. But they also made a chipset for the AthlonMP and it matters as much as the Opeteron one does. What I am getting at is they are both server cpus that were supported.

    Where are the chipsets for the T-bird, Athlon XP, Duron, Athlon 64, Athlon FX, and now Sempron.
  • coldpower27 - Wednesday, July 28, 2004 - link

    AMD reliability, has gone up alot since it earlier days of the K6 Series, though the bad image of their lack of reliability takes time to shake off.

    Though I see Intel recalling things as a good sign, it means if it doesn't meet Intel expectation they are willing to do what it takes to fix the problem.
  • coldpower27 - Wednesday, July 28, 2004 - link


    Well I would consider the Sempron 3100+ a reasonable buy, you can get a good Nforce 250 mobo with it, and have an upgrade path to at least Athlon 64 3700+. There are also gonna be extensions to that line to the Sempron line on Socket 754 as we go along.

    Though It's a sorta in between Socket though, more advanced then Socket 462 but lower then 939.

    LGA775 from Intel is probably best compared to Socekt 939. There will be LGA Celeron D's release in the future of course, when they will be offering, a suitable upgrade path with the I915 chipsets.

    The Sempron for Socket A is a meh, I would say a Sempron 2800+ just edges out the Celeron D.

    Is the Sempron 2800+ at least lower multiplier unlocked? Then you can simulate the other Sempron speeds from 2200+ to 2600+

    And there is no point making to try an apples to apples comparison with AMD and Intel, their architectures just differ to greatly. K7 and K8 both seem to rely on their LV1 cache quite well, as they cope well with reduce LV2 cache.

    Intel on the other hand has a small amount of LV1 Cache, and with Prescott their LV1 cache is not much faster then K8's or K7's is. not to mention that deep 31 stage pipeline. So 256Kb of minimum LV2 seems to be pretty critical.

    Personally to me I would take a Nforce 250GB with the Sempron 3100+ that is a very good choice for me.

    I wondered if your allowed to get the top off the Sempron 3100+ and see if it is a Newcastle with disabled LV2 cache or actually a Paris core with native 256KB of cache.

    As said I would also like to include the Pentium 4 2.4A for comparison purposes, as well as the lower Semprons, a Pentium 4 2.8C would be nice as well.

    Also with the inroduction of the Smepron AMD, can make more money, since these are only 84mm2 cores. Unless they decide to use some Thorton core's too.
    Dunno about the Sempron 3100+... They saw they were offering too much value in the low end space so they decided to reduce it. :D so they can earn more money.

    Of course an Mobile Barton 2400+ 35W will offer alot more value, but it was never AMD's intention of targetting the value sector with that processor :D

    We might see value still see more chips for Socket 754 since mobile chips are based on this socket.
  • Marlin1975 - Wednesday, July 28, 2004 - link

    AMD does make chipsets.

    They have the opteron chips for AGP, PCI-X, etc...
    And that is not even including that EVERY A64/FX/Opteron has the memory con. built in now.

Log in

Don't have an account? Sign up now